The effect of imputation of exposure estimates on the association between fine particulate matter and mortality

PURPOSE: The Harvard Six Cities Study (HSCS) found a small but significant association between daily PM2.5 and daily mortality count. The HSCS findings have been used as the basis for new EPA regulations, requiring lower levels of PM2.5. We feel that there are unresolved issues regarding the HSCS that should be fully evaluated prior to its findings being used as the basis of new regulation, including how the extent and method of imputing exposure data affect the association with daily mortality counts.METHODS: We examined the association between PM2.5 levels and daily mortality count, comparing the results from the HSCS methods with results based on an alternate imputation method, and with non-missing data.RESULTS: Overall, approximately 30% of the data points used in the HSCS were imputed. The method of imputation affected the association between particulate matter and mortality to a substantial degree in most of the cities. When the model using the HSCS method was compared to the model using the alternate method, in two areas the coefficients decreased substantially and lost significance. In two areas they changed little; in one area it rose substantially and became significant; and in one area it declined substantially but remained significant. When compared to the model based on the non-missing data, somewhat different patterns were observed. In both comparisons there were some large changes in the magnitude of the effect, but these were not consistent with the model used.CONCLUSIONS: We are concerned about the degree of data imputation and the effect that the method of imputation has on the association between particulate matter levels and mortality. In the case of the HSCS it appears that the imputed data are more strongly associated with the outcome than other methods of imputation and than the non-missing data. The reasons for these observations are not readily apparent, but the differences in effect should be explored and explained.